ZIPLORA

Summary from the TLDR AI newsletter

LoRAs are low rank learned matrices that change model behavior but are cheaper than regular fine-tuning. This work proposes an effective way to merge LoRAs together while maintaining the novel information from each.